1 STUN: SPATIO-TEMPORAL UNCERTAIN (SOCIAL) NETWORKS Chanhyun Kang Computer Science Dept. University of Maryland, USA chanhyun@cs.umd.edu Andrea Pugliese DEIS Dept. University of Calabria, Italy apugliese@deis.unical.it John Grant, V.S. Subrahmanian Computer Science Dept. University of Maryland, USA {grant,vs}@cs.umd.edu 2 Motivation Let’s assume that there is a social network including spatio-temporal information with certainty values. Maryland Potomac Bethesda 3 Motivation • Query example • Find all people who attended a party in Maryland at time point 5 with certainty at least 0.5 Common subgraph matching query Temporal constraint Certainty constraint Spatial constraint Within Maryland At time point 5 At least 0.5 certainty The query contains not only common graph query but also constraints for spatio-temporal information and certainty values 4 Motivation • In graph query research • Several subgraph matching algorithms and index structures are suggested • The indexes and the algorithms consider graph structure property only • But in order to answer the query efficiently, we need to consider • Graph structure property • Spatio-temporal information property • Certainty information property • So, we suggest a new index structure considering the properties and a query processing algorithm using the index. 5 In this paper • Introduce STUN: Spatio-Temporal Uncertainty (Social) Network • Define STUN query language • Develop STUN index, a disk based index structure • Develop a query processing algorithm using STUN index • Evaluate the algorithms 6 STUN • Spatio-Temporal Uncertainty (Social) Network is an extension of social networks • Supports aspects of spatio-temporal uncertainty in networks • Where and when the relationships are/were true • How certain we are that the relationships hold/held • Defined by a set of STUN tuples • STUN tuple : STUN quadruple + STUN annotation • STUN quadruple : two vertices, a relationship and a certainty value • STUN annotation : spatio-temporal information 7 Syntax : STUN quadruple • STUN quadruple : (v, l, v’; c) • v, v’ ∈ V (vertices) and l ∈ L (labels) • Certainty factor c ∈ [0,1] For example, (Jim, Friend, Ed; 0.7) Jim Ed Friend;0.7 “Jim” is a friend of “Ed” with certainty 0.7 8 Syntax : STUN annotation • STUN annotation: [R,T] • Expresses spatial information and temporal information • R is a region, a set of space points in a spatial reference system S • S ⊆ [0,M] x [0,N] with M,N ∈ R (Real numbers) • A space point is a member of S • T is a time interval, a pair(st, et) with st ≤ et • st and et are time points to express the start and the end of a specific period • A time point is a member of a temporal reference system [L, U] 9 Syntax : STUN tuple • STUN tuple : (v, l, v’; c) : [R, T] • STUN quadruple + STUN annotation Ex. (Phil, Organized, Party2; 1):[Bethesda, (15,15)] Party2 Phil ( ,Organized, ;1) [Bethesda, (15,15)] “Phil” organized “Party2” with certainty 1 and the event occurred at time 15 at some location within the region “Bethesda” • A STUN knowledge base is a finite set of STUN tuples. 10 STUN QUERY LANGUAGE 11 STUN Queries • A STUN query q contains • Graph part (Gq) • Subgraph query • Minimum certainty values for the relationships in the graph query • Constraint Part (Cq) • Constraints for spatial information • Constraints for temporal information Example. Find all people who attended a party in Maryland at time point 5 Subgraph query with certainty at least 0.5 Minimum certainty value Constraint for spatiotemporal information 12 STUN Queries • Graph part : Gq • Subgraph query and Minimum certainty values • A set of query graph tuples • Variables are denoted using “?”; output variables are underlined • A query graph tuple is (v, l, v’; c) : [R, T] where • v, v’∈ V U VARV, l ∈ L U VARL, c∈[0,1], • R ∈VARR and T ∈VART Example. Find all people(?I) who attended a party(?P) in Maryland at time point 5 Subgraph query with certainty at least 0.5 Gq={(?I, attended, ?P; 0.5):[?s,?t]} Minimum certainty value 13 STUN Queries • Constraint part: Cq • Specify spatial constraints and temporal constraints • Expressed by • Predicate symbols • Represent a spatial relation or a temporal relation • Parameters for the predicates • Ground terms or variables in the graph part Example. Find all people(?I) who attended a party(?P) in Maryland at time point 5 with certainty at least 0.5 Spatial constraint Temporal constraint Cq ={inside(?s, Maryland), during(?t,[5,5])} 14 STUN Query example • Find all people(?I) who attended a party(?P) in Maryland at time point 5 with certainty at least 0.5 Gq={(?I, attended, ?P; 0.5):[?s,?t]} Cq ={inside(?s,Maryland), during(?t,[5,5])} 15 STUN Query example • Finds all people(?I) • who have been a friend of ‘Jim’ in the time interval [10,20] with certainty at least 0.9 as well as a friend of ‘Phil’ in the same interval with certainty at least 0.6 • And who attended a party(?P) in Maryland organized by ‘Phil’ that occurred during the time interval [0,20] Gq={(?I, attended, ?P; 1.0):[?s1,?t1], (?I, friend, Jim; 0.9):[?s2,?t2], (?I, friend, Phil; 0.6):[?s2,?t2], (Phil, organized, ?P; 1.0):[?s1,?t1],} Cq={inside(?s1, Maryland), during(?t1,[0,20]), during(?t2,[10,20])} 16 STUN query answer • A substitution θ maps variables to ground terms • Each ground term maps to itself • Denote the application of θ to a term x as xθ Phil Organized ?P Phil Organized Party3 Substitution θ • A substitution θ is an answer to a STUN query q:(Gq, Cq) • The tuples with θ for the Gq exist in the STUN KB • The certainty values of the tuples in STUN KB are larger than minimum certainty in the Gq • The spatio-temporal information of the tuples satisfy all constraints in the Cq - ∀ π£, π, π£ ′ ; π : π , π ∈ πΊπ, ∃π ′ ≥ π π . π‘. π£π, ππ, π£ ′ π; π ′ : π π, ππ ∈ πΎπ΅ - And ππ∈πΆπ πΆππ is true 17 STUN INDEX 18 STUN Index • A balanced tree • Each leaf node represents a portion of the STUN knowledge base. • Each inner node captures the subgraph represented by its child nodes. 19 STUN Index • Each node occupies a disk page and contains • MBR(minimum bounding rectangle) • Envelops the regions associated with the STUN tuples in the subgraph of child nodes • MBI(minimum bounding interval) • Envelops the time intervals associated with the STUN tuples in the subgraph of child nodes A spatial reference system R1,R2, R2: regions N1,N2,N3: nodes N3 MBRMBIs of N3 are used to prune nodes for • On processing queries, MBRs and the answers using R3 spatial constraints and temporal constraints R1 R2 N2 N1 MBR of N1 MBR of N2 R1 R2 R3 20 STUN Index • Reduce the number of nodes to read for answering queries. • Each index node should have • Few cross edges with other nodes at the same level • Small MBR(minimum bounding rectangle) and small MBI(minimum bounding interval) • Small MBR overlaps with other nodes at the same level • Small MBI overlaps with other nodes at the same level. • In order to achieve the constraints • Build a vertex and edge weighted undirected graph(WUG) from the STUN KB • Then, handle the weights on building the index 21 Building STUN Index I. Initial step • Build a vertex and edge weighted undirected graph(WUG) from STUN KB • The weights are used to satisfy the constraints • Few cross edges • Small MBR(minimum bounding rectangle) and small MBI(Minimum bounding interval) • Small MBR overlaps and small MBI overlaps II. Coarsening Step • Merging vertices using weights of vertices and edges III. Partitioning Step • Build a tree index using coarsened graphs 22 Building Index- Initial Step Initial Step I. • Build a vertex and edge weighted undirected graph(WUG) • Assign weights of vertices as 1 • Calculate weights of edges using a spatio-temporal vertex distance function πΏ π£, π£ ′ • Calculate MBR(minimum bounding rectangle)s and MBI(minimum bounding interval)s for edges e1 v0 v1 v2 e0 e2 Each edge contains a spatio-temporal information with a certainty value v0 1 MBR ππ, ππ MBR ππ v1 MBI ππ, ππ v2 MBI ππ 1 1 πΉ ππ, ππ πΉ ππ, ππ {πππ, πππ} {πππ} labels WUG 23 Building Index- Initial Step ο§ Spatio-temporal vertex distance function πΏ π£, π£ ′ • Looks at the neighborhood of the two vertices • Measures the “amount” of space and time the vertices share with each other with respect to their neighborhoods. πΏ π£, π£ ′ = πΌ β ππ π£, π£ ′ β ππ π£, π£ ′ = ππ π£ = ππ π£, π£′ = ππ π£ = 1 ππ π£ + π£,π,π£ ′ ;π :[π ,π]∈πΎπ΅ π π£,π,π£ ′ ;π :[π ,π]∈πΎπ΅ π πΌ + π½ = 1, πππ πΌ, π½ ∈ π β ππππβ π , β ππππ π , β ππππ π , 1 ππ π£ + π½ β ππ(π£, π£′)( β ππππβ π , π£,π,π£ ′ ;π :[π ,π]∈πΎπ΅ π π£,π,π£ ′ ;π :[π ,π]∈πΎπ΅ π 1 ππ π£ ′ + 1 ππ π£ ′ ), 24 Building Index- Coarsening • Coarsen the graph until the size of the coarsened graph is less than 1 disk page • At each coarsening level l, the number of vertices in Gl is half of the number of vertices in Gl-1 Coarsening Level k … … Level 2 N/2k Gk Merging vertices G2 N/4 Merging vertices Level 1 Level 0 G1 N/2 Merging vertices N G0 Original graph # of vertices 25 How to merge vertices Choose a vertex v randomly to merge Select a neighbor m of v with minimum edge weight (v is merged into m) Update the weight of vertex m : π π = π π + π π Update the weight, MBR and MBI of edges of v and m (If there is no edge between m and a neighbor of v, add an edge between m and the neighbor) Delete the edge between v and m and the vertex v Update mapping information: π(π −π (v)) 26 Building Index- Partitioning - Each edge already has a MBR and a MBI 1. Store Gk as a root page MBR(all edges of Gk) MBI(all edges of Gk) Gk 2. Partition 3.Induce subgraphs using the mapping information from Gk-1 to Gk 4. Store the subgraphs as child pages Gk-1 MBR(all edges of a) MBI(all edges of a) b a MBR(all edges of b) MBI(all edges of b) Gk-2 … … Coarsened graphs 5. Do the works until at the lowest coarsening level recursively 27 Query Answering • STUN index is used to get candidates for variables • Retrieve the index tree using mapping information with ground terms(constants) in a query • MBR(minimum bounding rectangle) and MBI(minimum bounding interval) are used to filter out the unnecessary pages for the query answer with regard to spatial and temporal constraints ?I friend Jim ?I friend Phil Phil organized ?P - Check MBRs and MBIs of pages with the constraints for pruning STUN index 28 Query Answering • Overall algorithm I. Get candidates for each variable of a query II. Select a variable that has the smallest number of candidates III. Substitute each candidate for the variable IV. For each substitution, do steps II and III for remaining variables recursively in a depth first manner V. If no variable is left, return the substitutions 29 EVALUATION 30 Experiment : Environment • We developed a prototype implementation in about 10,600 lines of Java code • Ran the code on a laptop • a dual-core 2.8 GHz CPU with 8G of RAM running Window 7 • Indexes are on the disk (No explicit buffer to load the index) • Experiments for the scalability of the STUN index by varying • The size of the graph • The complexity of queries • The number of constraints in queries • Queries are randomly generated from STUN KBs • Each query has at least one answer. • More than 10000 queries are tested 31 Experiment : Dataset • YouTube dataset • Vertices : people and groups • 20% of groups have a region randomly assigned • Edge relations • ‘follow’ : person to person, a time interval • ‘membership’ : person to group, a time interval • ‘co-located’ : person to group, a time interval and a region • Time intervals are randomly assigned to ‘follow’ and ‘membership’ relationships • A ‘co-located’ edge is added between two members if • They have ‘membership’ relationships with a same group • And they have overlapped time interval with the same group • And the same group has an assigned region 32 Experiment: Result • Every single data point was obtained by running 200 queries. 33 Experiment: Result • The query processing time increases slightly super-linearly with the size of the database thought the slope of the graph increases with the complexity of the query. 34 Conclusion • Introduce Spatio-Temporal Uncertainty (Social) network • Define STUN query language • Develop a disk based index structure • Develop a query processing algorithm • Do experiments for evaluating the STUN system 35 Questions